AI in a Changing Ocean: Big Data Analytics for Modeling Physical and Biogeochemical Ocean Dynamics

The dynamics of the oceans and marine ecosystems play a fundamental role in the Earth sustaining global biogeochemical cycles, regulating climate, and supporting life. Global changes, in particular climate change and biodiversity loss, are affecting these dynamics, driving shifts in the physical and ecological systems at an alarming rate. The systems and dynamics regulating the ocean are not fully understood and generally underexplored, with non-linear interactions that can generate more anomalies and extreme events having major impacts on the conditions of human life. Large oscillations can effectively change the dynamics and the spatiotemporal distribution of ocean variables into new stable systems, so-called tipping points.

Understanding how these systems function, how they are changing, and what drives those changes, is essential for anticipating future dynamics and providing early warning signals for policymakers and stakeholders. Achieving such understanding generally requires both large volumes of high-resolution data and the application of novel analytical methods. Specifically, it involves: (1) mapping ocean variables to specific spatial, vertical, and temporal domains; (2) learning how system dynamics evolve and identifying their key drivers; and (3) estimating how these variables may shift under future scenarios, including changes in the magnitude or direction of their influences.

In this thesis, state-of-the-art artificial intelligence algorithms based on neural networks are employed to extract patterns and relationships from various data sources, including satellite observations, in situ measurements, and numerical model outputs, with the aim of improving our understanding of the ocean using existing data. In parallel with developing predictive models, the present work examines their internal logic to improve understanding of model behaviour and improve confidence in their outputs.

One central focus of the thesis is the reconstruction of subsurface ocean structure, specifically temperature and salinity profiles, using only surface data and geospatial information. A convolutional neural network is trained to infer full vertical profiles from satellite-derived fields. Applied in regions including the Atlantic, East Greenland, and the Black Sea, the model demonstrates strong performance across varying conditions, offering a scalable approach for reconstructing the full three-dimensional structure of the ocean column, even in regions with limited in situ observations.

A second major component examines the dynamics of chlorophyll, a proxy for phytoplankton biomass, in the Black Sea. Using output from a coupled physical–biogeochemical model as training data, a 3D convolutional sequence-to-sequence network is developed to emulate chlorophyll evolution over time. This emulator not only replicates the behavior of the original numerical model with high accuracy but also provides insights into the relative importance of different physical and biogeochemical drivers across space and season. When applied to future projections under a high-emission climate scenario, the model captures shifts in dominant drivers, offering a data-driven perspective on how ecosystem dynamics may evolve in the coming decades.

In both projects, the analysis of relative importance and sensitivity of input variables played a central role. For the reconstruction task, the influence of each surface variable was assessed across depth layers to identify which surface features most strongly influence subsurface temperature and salinity at different depths. In the chlorophyll emulator, importance analyses were used to disentangle seasonal and regional drivers of variability, offering insight into how ecosystem dynamics may shift under changing environmental conditions.

These analyses demonstrate how deep learning methodologies can play a role in addressing fundamental challenges in ocean science, by considerably improving our ability to reconstruct essential physical ocean properties and accurately emulate complex non-linear ecological dynamics. The reconstruction of ocean temperature and salinity profiles is crucial, as these properties sustain the stability and functioning of marine ecosystems, directly influencing circulation patterns, nutrient availability, and habitat suitability. Understanding and accurately predicting chlorophyll dynamics complements this by providing insight into ecosystem productivity, health, and resilience. Together, these advancements address the key objectives of this thesis, mapping ocean variables, uncovering system dynamics and drivers, and projecting future changes, and may be used to support earlier detection of changes in the ocean offering practical information that can help policymakers and stakeholders to respond to climate impacts, protect marine biodiversity, and ensure sustainable use of ocean resources.